Inference SDK v0.1.3

OpenAI-compatible speech engines now honor the request's voice over the engine's configured default; the response reports the voice actually sent. Engine defaults still apply when no voice is requested. Covers Kokoro and other kind: "openai" TTS adapters.

Assets: inference-sdk-0.1.3.tar.gz (28 files, manifest digest 1353deb0f289) and SHA256SUMS. Validate with pnpm-docker run typecheck && pnpm-docker run test — 61 tests.

Inference SDK v0.1.2

The first public compiled release of @bitdeep/inference-sdk.

The SDK coordinates LLM, ASR and TTS requests, demand loading, idle unload and model handoffs through one process-local GPU owner. It includes adapters for vLLM, Speaches/Whisper, Chatterbox, Qwen3, Kokoro and Text Embeddings Inference.

Version 0.1.2 adds the public distribution metadata and architecture guide to the existing implementation. Runtime behavior is unchanged from 0.1.1.

Download

1. Download inference-sdk-0.1.2.tar.gz and SHA256SUMS.
2. Run sha256sum --check SHA256SUMS.
3. Extract the archive and verify its files against MANIFEST.json.
4. Install or vendor the extracted directory through your application's dependency import process, pinning its version and manifest hash.

The archive contains compiled JavaScript, TypeScript declarations, package metadata, documentation and licensing. It has no third-party runtime dependencies and includes no model weights, customer data or deployment configuration. The package is not published to npm.

Node 26 is required. FFmpeg is required in the worker image for MP3 conversion and multi-part speech.

Validation: typecheck, 60 unit tests and the compiled build passed for this release. The unchanged engine implementation previously passed real FFmpeg checks and a synthetic GPU flow across chat and speech, including ASR reload. These are functional checks, not performance benchmarks.

Read the quickstart and architecture guide before sharing a GPU. Authentication, queues and customer isolation remain the application's responsibility.